Hi, high-poly shouldn't matter if the UVs of the mirrored parts are offset. It will bake for UVs inside 0-1 space, mirrored parts with UVs offset by 1 will reuse the baked information: high-poly details are mirrored.
When a cage is used the lowpoly vertex normals are ignored for projection and instead vertex will point towards their twin on the cage model. Since we want them all looking at the same tiny point in space we just shove the entire cage into that same spot. Scaling to 1% instead of 0% avoids bits of the cage getting merged…